Identifying bioactivity events of small molecules from the scientific literature

The PhD work focused on the automatic extraction of small molecule bioactivity information from scientific literature, a topic that ultimately aims to contribute to the small molecule ontology ChEBI. The ChEBI ontology plays an important role in assisting biomedical research, but requires further expansion in order to improve the ability to interrelate and link ChEBI to other primary biological resources. Small molecule bioactivity is a concept that presents small molecule characteristics in scenarios involving other biological objects. Bioactivity effects can have various targets including genes, proteins, organs, and complete organisms. Two methods of extracting these events were examined. In the first, a supervised machine learning approach was proposed and developed based on a new corpus: DrugPro. The developed system comprised cascaded classifiers and a post-processing unit to extract the specific relationship between drugs and bacteria. The classification system delivered a good result when combining domain-specific features and a term weighting scheme, achieving precision of 92.56% and recall of 93.22% on the test corpus. A sentence classification system also achieved reasonable results, with precision of 58.9% and recall of 57.4% on the test corpus. A small set of language patterns was developed to extract the key items of information from the positive sentences. The second method of extracting small molecule bioactivity events involved the development of an entirely rule-based extraction system. The performance of the system was evaluated on a random selection of 20 abstracts, which resulted in 94.7% precision and 94.7% recall on this (admittedly limited) data set. The results indicated the rule-based approach has the potential to precisely identify general small molecule bioactivity events in a much larger collection of biomedical text. It is concluded that extension and enhancement of the ChEBI ontology could be based on the concept of small molecule bioactivity, and text mining provides a viable strategy by which this can be implemented. A further contribution of the work is the set of text mining resources created here, including the annotated corpus, and an investigation into the best way to undertake named-entity recognition for the multi-domain concept of bioactivity. It